Bark

Bark won on emotional, lifelike sound, not on standard speech metrics. Bark made some word errors, yet it sounded far more emotional and lifelike than text-to-speech systems tuned for word error rate and speed. It could even whisper, which no speech system did.

Earlier text-to-speech sounded robotic and never emotional. Mikey Shulman's view is that audio is "so much more subjective, so much more gnarly, and so much more beautiful than text," and that asking "how do I do the same application faster" misses what is only possible today.

Subversive▶ 15:30

Open-sourcing an unprofitable model validated demand before Suno commercialized it. Bark was too costly and inefficient to run commercially, so Suno released it as open source. It got a lot of GitHub stars very quickly and a developer community formed around it.

The reaction showed demand for expressive audio beyond reading text aloud and was the spark that confirmed there was something in audio. Bark was the first breakthrough toward passable audio for the team.

Subversive▶ 12:58

Suno used Bark's lesson to pick music over the bigger-seeming speech market. After Bark, many people told the team to build a speech company because it looked bigger and more straightforward. Suno chose music, seeing more value in artistic uses than in reading text aloud.

At the time everyone wanted text read aloud and ignored artistic uses. Shulman notes that speech players like ElevenLabs and Speak compete on accuracy, voice variety and coaching, not on music.

Subversive▶ 15:30